liveBook · Manning

Taste drift

As explained in AI Recommender Systems

Taste drift is the gradual or sudden change in a user's preferences over time.

Taste drift is the gradual or sudden change in a user's preferences over time. Historical behavior can therefore become a stale guide: a user who favored superhero movies months or years ago may no longer want them today. section 4.1.1 section 4.7

Account for taste drift by:

  • Limiting training data to a recent time window so current preferences receive more weight. section 4.7
  • Blending a recent-history model with a long-term-history model to balance current interests and durable preferences. section 4.7
  • Monitoring declining engagement with recommendations as a possible signal that the system is relying on outdated preferences. section 4.7
  • Evaluate user segments separately, because performance for long-term users can hide failures elsewhere in the lifecycle. section 4.7

A stale preference profile

A user who loved superhero movies three years ago may have moved on. A model trained on the full history can keep recommending them, making the experience feel stale; recent-history weighting or model blending can respond to the change. section 4.7

Figure 4.19 The different stages of a user's lifecycle, and which types of recommendations can be used at each stage.

Use the lifecycle figure to place taste drift alongside cold-start, active-user, and long-term-user recommendation strategies. section 4.7

How AI Recommender Systems covers it

How deep it goes
Defines temporal preference change and gives mitigations, but stops short of specifying model architectures or adaptation algorithms.
Examples
Uses changing superhero-movie interest and declining recommendation engagement to illustrate stale profiles and detection signals.
What it assumes
Assumes recommendation histories and user lifecycle segments, while treating changing preferences as a practical evaluation concern.
What you can do after
Apply it via recent-history windows, blended models, and engagement monitoring to reduce stale recommendations.
How it connects
Connects taste drift with recommendation systems, offline evaluation, collaborative filtering, content-based filtering, and user lifecycle analysis.

This explainer was generated by Manning's knowledge graph from the text of AI Recommender Systems. The section links point to the parts of the book it draws on.